biological modeling of neural networks

Biological Modelingof Neural Networks

Week 5

NETWORKS of NEURONS and

ASSOCIATIVE MEMORY

Wulfram GerstnerEPFL, Lausanne, Switzerland

5.1 Introduction - networks of neuron - systems for computing - associative memory

5.2 Classification by similarity

5.3 Detour: Magnetic Materials

5.4 Hopfield Model

5.5 Learning of Associations

5.6 Storage Capacity

Week 5: Networks of Neurons-Introduction

Systems for computing and information processing

Brain Computer

CPUmemory

input

Von Neumann architecture

(10 transistors)1 CPU

Distributed architecture10

(10 proc. Elements/neurons)No separation of processing and memory

10

10 000 neurons3 km wire

1mm

Systems for computing and information processing

Brain

Distributed architecture10

10 neurons

No separation of processing and memory

4 10 connections/neurons

Systems for computing and information processing

10 000 neurons3 km wire

1mm

Associations, Associative Memory

Read this text NOW!

I find it rea*l* amazin* t*at y*u ar* abl* to re*d t*is tex* despit* th* fac* *hat more t*an t*ent* perc*n* of t** char*cte*s a*e mis*ing.*his mean* t*at you* brai* i* abl* ** fill in missin* info*matio*.

Noisy word

pattern completion/word recognition

brain

List of words

brai*

atombravebrainbrass

Output the closest one

Your brain fills in missing information: ‘associative memory’

Biological Modelingof Neural Networks

Week 5

NETWORKS of NEURONS and

ASSOCIATIVE MEMORY

Wulfram GerstnerEPFL, Lausanne, Switzerland

5.1 Introduction - networks of neuron - systems for computing - associative memory

5.2 Classification by similarity

5.3 Detour: Magnetic Materials

5.4 Hopfield Model

5.5 Learning of Associations

5.6 Storage Capacity

Week 5: Networks of Neurons-Introduction

image

5.2 Classification by similarity: pattern recognition

Classification:comparisonwith prototypes

AB

T

Z

PrototypesNoisy image

PrototypesNoisy image

Classification by closest prototype

AT pxpx Blackboard:

5.2 Classification by similarity: pattern recognition

Noisy image

Aim: Understand Associative MemoryAssociative memory/collective computation

Brain-style computation

Full image

Partial word

Full word

5.2 pattern recognition and Pattern completion

Quiz 5.1: Connectivity

A typical neuron in the brain makes connections-To 6-20 neighbors-To 100-200 neurons nearby-To more than 1000 neurons nearby-To more than 1000 neurons nearby or far away.

In a typical cristal in nature, each atom interacts-with 6-20 neighbors-with 100-200 neurons nearby-with more than 1000 neurons nearby-with more than 1000 neurons nearby or far away.

Biological Modelingof Neural Networks

Week 5

NETWORKS of NEURONS and

ASSOCIATIVE MEMORY

Wulfram GerstnerEPFL, Lausanne, Switzerland

5.1 Introduction - networks of neuron - systems for computing - associative memory

5.2 Classification by similarity

5.3 Detour: Magnetic Materials

5.4 Hopfield Model

5.5 Learning of Associations

5.6 Storage Capacity

Week 5: Networks of Neurons-Introduction

5.3 Detour: magnetism

S

N

Noisy magnet pure magnet

5.3 Detour: magnetism

Elementary magnet

Si = +1

Si = -1 Blackboard: example

5.3 Detour: magnetism

dynamics

Sum over all interactions with i

( 1) sgn[ ( )]i jj

S t S t

Elementary magnet

Si = +1

Si = -1

blackboard

Anti-ferromagnet

wij = +1

wij = -1

5.3 Detour: magnetism

dynamics

Sum over all interactions with i

( 1) sgn[ ( )]i ij jj

S t w S t

5.3 Magnetism and memory patterns

Elementary pixel

Si = +1

Si = -1

blackboardwij = +1

wij = -1

wij = +1

Hopfield model:Several patterns next section

dynamics

Sum over all interactions with i

( 1) sgn[ ( )]i ij jj

S t w S t

Exercise 1: Associative memory (1 pattern)

Elementary pixel

Si = +1

Si = -1

wij = +1wij = +1

9 neurons - define appropriate weights - what happens if one neuron wrong? - what happens if n neurons wrong?

Next lecture at 10h15

dynamics

Sum over all interactions with i

( 1) sgn[ ( )]i ij jj

S t w S t

Biological Modelingof Neural Networks

Week 5

NETWORKS of NEURONS and

ASSOCIATIVE MEMORY

Wulfram GerstnerEPFL, Lausanne, Switzerland

5.1 Introduction - networks of neuron - systems for computing - associative memory

5.2 Classification by similarity

5.3 Detour: Magnetic Materials

5.4 Hopfield Model

5.5 Learning of Associations

5.6 Storage Capacity

Week 5: Networks of Neurons-Introduction

5.4 Hopfield Model of Associative Memory

dynamics

Sum over all interactions with i

Hopfield model

Prototype

p1

Prototype

p2

jiij ppw

interactions

Sum over allprototypes

( 1) sgn[ ( )]i ij jj

S t w S t

dynamics

all interactions with i

DEMO

Random patterns, fully connected: Hopfield model

Prototype

p1

jiij ppw

interactions

Sum over allprototypes

This ruleis very goodfor randompatterns

It does not work wellfor correlated patters

5.4 Hopfield Model of Associative Memory

( 1) sgn[ ( )]i ij jj

S t w S t

5.4 Hopfield Model of Associative Memory

jiij ppw

1( ) j jm t S t

N overlap

1( 1) 1j jm t S t

N

Blackboard

( 1) sgn[ ( )]i ij jj

S t w S t

Hopfield model

Prototype

p1 Finds the closest prototypei.e. maximal overlap (similarity)

m

Interacting neurons

Computation - without CPU,- without explicit memory unit

5.4 Hopfield Model of Associative Memory

Exercise 3 (now)

Prototype

p1

Assume 4 patterns. At time t=0, overlap withPattern 3, no overlap with other patterns. discuss temporal evolution (of overlaps) (assume that patterns are orthogonal)

1ij i jNw p p

Sum over all interactions with i

Next lecture at 11h15

( 1) sgn[ ( )]i ij jj

S t w S t

Biological Modelingof Neural Networks

Week 5

NETWORKS of NEURONS and

ASSOCIATIVE MEMORY

Wulfram GerstnerEPFL, Lausanne, Switzerland

5.1 Introduction - networks of neuron - systems for computing - associative memory

5.2 Classification by similarity

5.3 Detour: Magnetic Materials

5.4 Hopfield Model

5.5 Learning of Associations

5.6 Storage Capacity

Week 5-5: Learning of Associations

Hebbian Learning

pre jpost

iijw

When an axon of cell j repeatedly or persistently takes part in firing cell i, then j’s efficiency as oneof the cells firing i is increased

Hebb, 1949

k

- local rule- simultaneously active (correlations)

Where do the connections come from?

5.5 Learning of Associations

5.5 Hebbian Learning of Associations

item memorized

5.5 Hebbian Learning of Associations

item recalled

Recall:Partial info

5.5 Hebbian Learning: Associative Recall

5.5 Associative RecallTell me the object shape for the following list of 5 items:Tell me the color for the following list of 5 items:

be as fast as possible:

time

Tell me the color for the following list of 5 items:

RedBlueYellowGreenRed

Stroop effect:Slow response: hard to work Against natural associations

be as fast as possible:

time

5.5 Associative Recall

Hierarchical organization of Associative memory

animals

birds fish

Name as fast as possiblean example of a bird

swan (or goose or raven or …)

Write down first letter: s for swan or r for raven …

5.5 Associative Recall

name as fast as possiblean example of a

hammer

red

Apple

violin

tool

color

music instrument

fruit

Nommez au plus vite possibleun exemple d’un /d’une

outilcouleur

fruitinstrumentde musique

5.5 Associative Recall

Biological Modelingof Neural Networks

Week 5

NETWORKS of NEURONS and

ASSOCIATIVE MEMORY

Wulfram GerstnerEPFL, Lausanne, Switzerland

5.1 Introduction - networks of neuron - systems for computing - associative memory

5.2 Classification by similarity

5.3 Detour: Magnetic Materials

5.4 Hopfield Model

5.5 Learning of Associations

5.6 Storage Capacity

Week 5-5: Learning of Associations

learning of prototypes

Prototype

p1

Prototype

p2

interactions

Sum over allprototypes

(1)

Q; How many prototypes can be stored?

dynamics

all interactions with i

jiNij ppw 1

( 1) sgn[ ( )]i ij jj

S t w S t

Prototype

p1

Prototype

p2

jiij ppw Interactions (1)

Q; How many prototypes can be stored?

Dynamics (2)

Random patterns

Minimal condition: pattern is fixed point of dynamics-Assume we start directly in one pattern-Pattern stays

Attention: Retrieval requires more (pattern completion)

blackboard

( 1) sgn[ ( )]i ij jj

S t w S t

Exercise 4 now: Associative memoryQ; How many prototypes can be stored?

Random patterns random walka) show relation to erf function: importance of p/Nb) network of 1000 neurons – allow at most 1 wrong pixel?c) network of N neurons – at most 1 promille wrong pixels?

End of lecture, exercise+ Computer exercise : 12:00

Prototype

p1

Prototype

p2

jiij ppw Interactions (1)

Dynamics (2) ( 1) sgn[ ( )]i ij jj

S t w S t

Prototype

p1

Prototype

p2

jiij ppw Interactions (1)

Dynamics (2)

Random patterns

Minimal condition: pattern is fixed point of dynamics-Assume we start directly in one pattern-Pattern stays

Attention: Retrieval requires more (pattern completion)

Week 6 Review: storage capacity of Hopfield model

( 1) sgn[ ( )]i ij jj

S t w S t

10

errorP

Q; How many prototypes can be stored?

Biological Modelingof Neural Networks

Week 6

Hebbian LEARNING and

ASSOCIATIVE MEMORY

Wulfram GerstnerEPFL, Lausanne, Switzerland

Week 6: Hopfield model continued

6.1 Stochastic Hopfield Model6.2. Energy landscape6.3. Low-activity patterns6.4. Attractor memorie - spiking neurons - experimental data

Prototype

p1

Prototype

p2

interactions

Sum over allprototypes

(1)

Deterministic dynamics

jiNij ppw 1

6.1 Review: Hopfield model

( 1) sgn[ ( )]i ij jj

S t w S t

Prototype

p1

Prototype

p2

jiij ppw Interactions (1)

Dynamics (2)

Pr{ ( 1) 1| } [ ]i i i j ij jS t h g h g w S t

Random patterns

blackboard Pr{ ( 1) 1| }i i iS t h g p m t

6.1 Stochastic Hopfield model

( )m t t

( )m t

0( )m t

6.1 Stochastic Hopfield model: memory retrieval

3 1m 17 1m

6.1 Stochastic Hopfield model: memory retrieval

Dynamics (2)

Pr{ ( 1) 1| } [ ]i i i j ij jS t h g h g w S t

blackboard Pr{ ( 1) 1| }i i iS t h g p m t

6.1 Stochastic Hopfield model

Assume that there is only overlap with pattern 17: two groups of neurons: those that should be ‘on’ and ‘off’

17Pr{ ( 1) 1| }i iS t h h g m t 17Pr{ ( 1) 1| }i iS t h h g m t

17 17 17 17 172 ( 1) {1 } {1 }m t g m t g m t g m t g m t

( )m t t

( )m t

0( )m t

6.1 Stochastic Hopfield model: memory retrieval 17 17 17 17 172 ( 1) {1 } {1 }m t g m t g m t g m t g m t 17 17( 1)m t F m t

Biological Modelingof Neural Networks

Week 6

Hebbian LEARNING and

ASSOCIATIVE MEMORY

Wulfram GerstnerEPFL, Lausanne, Switzerland

Week 6: Hopfield model continued

6.1 Stochastic Hopfield Model6.2. Energy landscape6.3. Low-activity patterns6.4. Attractor memorie - spiking neurons - experimental data

3 1m 17 1m

6.2 memory retrieval

E

3 1m 17 1m

6.2 Symmetric interactions: Energy picture

Exercise 2 now: Energy pictureNext lecture11:25

E

3 1m 17 1m

ij i ji j

E w S S

( 1) sgn[ ( )]i ij jj

S t w S t

Biological Modelingof Neural Networks

Week 6

Hebbian LEARNING and

ASSOCIATIVE MEMORY

Wulfram GerstnerEPFL, Lausanne, Switzerland

Week 6: Hopfield model continued

6.1 Stochastic Hopfield Model6.2. Energy landscape6.3. Low-activity patterns6.4. Attractor memories - spiking neurons - experimental data

3 1m 17 1m

6.3 Attractor memory

Memory with spiking neurons-Mean activity of patterns?-Separation of excitation and inhibition?-Modeling?-Neural data?

6.3 attractor memory with spiking neurons

1 2 … Ni …

m=1

m=2

m=3

jiNij ppw 1

Random patterns +/-1 with zero mean 50 percent of neurons should be active in each pattern

6.3 attractor memory with low activity patterns

1 2 … Ni …

m=1

m=2

m=3

1 ( )( )ij i jNw p b p a

Random patterns +/-1 with low activity (mean =a<0) 20 percent of neurons should be active in each pattern

activitySome constant

Biological Modelingof Neural Networks

Week 6

Hebbian LEARNING and

ASSOCIATIVE MEMORY

Wulfram GerstnerEPFL, Lausanne, Switzerland

Week 6: Hopfield model continued

6.1 Stochastic Hopfield Model6.2. Energy landscape6.3. Low-activity patterns6.4. Attractor memories - spiking neurons - experimental data

Inh1

Inh2

theta

Exc

1 ( 1)( 1)ij i jNw p p

Inh1

Inh2Hebb-rule: Active together

6.4 attractor memory with spiking neurons

Spike raster

Overlap with patterns 1 … 3

overlaps

Spike raster

Overlap with patterns 1 … 11 (80 patterns stored!)

Memory with spiking neurons-Low activity of patterns?-Separation of excitation and inhibition?-Modeling?

-Neural data?

All possible

Sidney opera

Sidney opera

Sidney opera

Human Hippocampus

Quiroga, R. Q., Reddy, L., Kreiman, G., Koch, C., and Fried, I. (2005). Invariant visual representation by single neurons in the human brain. Nature, 435:1102-1107.

6.4 memory data

Delayed Matching to Sample Task

1ssample match

1ssample match

Animal experiments

6.4 memory data

20

1ssample match

[Hz]

Miyashita, Y. (1988). Neuronal correlate of visual associative long-term memory in the primate temporal cortex. Nature, 335:817-820.

6.4 memory data

match

20

[Hz]

sample

0 1650ms

0

Rainer and Miller (2002). Timecourse of object-related neural activity in the primateprefrontal cortex during a short-term memory task. Europ. J. Neurosci., 15:1244-1254.

6.4 memory data

In the Hopfield model, neurons are characterized by a binary variable Si = +/-1. For an interpretation in terms of spikes it is, however, more appealing to work with a binary variable xi which is zero or 1.(i) Write Si = 2xi- 1 and rewrite the Hopfield model in terms of xi.What are the conditions so that the input potential is

(ii) Repeat the same calculation for low-activity patterns and and weights

with some constants a and b

1 ( )( )ij i jNw p b p a

i j ij jh w x

Exercise 3 NOW- from Hopfield to spikes

The end